Future AGI

Customer support agent development

Build and evaluate support bots that use knowledge base retrieval to resolve issues step-by-step.

Agent performance benchmarking

Run evaluation runs to measure factuality, relevance, safety, and completeness of agent responses.

Simulated scenario testing

Create and test agents across complex, real-world interactions like debt collection with multiple conversation branches.

Safety and compliance testing

Implement global prompts to handle sensitive situations, such as suicide threats, hostile callers, or requests to speak with a human.

Iterative improvement

Compare agent versions (e.g., v1 vs. v2) to see performance gains and identify specific areas for optimization.

Self-improving agents

Build agents that automatically catch failures and update their behavior based on evaluation results.

Evaluation runs

Run tests that score agents on factuality, relevance, safety, and completeness, with detailed pass/fail results.

Scenario simulation

Create and edit simulated conversations with customizable personas, situations, and outcomes to test agent behavior under pressure.

Global prompt management

Define system-level prompts that trigger automatically for critical situations like suicide threats, hostile callers, or human transfer requests.

Version comparison

Compare agent versions side-by-side (e.g., v1 at 67% overall vs. v2 at 91%) to track improvement over time.

Knowledge base integration

Connect agents to vector retrieval tools for searching top-k relevant articles to ground responses.

Open-source flexibility

Licensed under Apache 2.0 with 986 GitHub stars, allowing full customization and self-hosting.